GEO Joint Experiment for Crop Assessment and Monitoring (JECAM): Template for Research Progress Report

Transcription

1 GEO Joint Experiment for Crop Assessment and Monitoring (JECAM): Date: 17/02/2015 JECAM Test Site Name: Brazil São Paulo Template for Research Progress Report Team Leader and Members: Guerric le Maire, CIRAD Yann Nouvellon, CIRAD ; Jean-Paul Laclau, CIRAD ; José-Luiz Stape, IPEF and UNESP ; Stéphane Dupuy, CIRAD Use of Information In addition to the report we would also like to use the information and images you provide to update the jecam.org website. Do you agree to this use of your information? Y, please ask if you need better quality images Project Objectives Have the original objectives for your site changed? Y/N Please briefly describe the areas that you are working on from the list below (i.e. topic, general methods, intended operational outcome, if any): Crop identification and Crop Area Estimation Site Description Location: Lat , Lon Topography: slope <5% in centroid area Soils: Ferralsols, 20% Clay (in centroid area) Drainage class/irrigation: Moderately to well drained, high water consumption for Eucalyptus stands, cropland sometimes irrigated Crop calendar: Eucalyptus: 6 years rotations ; Other crops and sugarcane: monitoring started in December 2014, but mainly sugarcane monoculture, oran Field size: 40 ha for Eucalyptus field, large fields for other crop classes Climate and weather: Humid Tropical (Aw Koppen), weather stations Agricultural methods used

2 Photograph(s) Many photographs have been taken during the December 2014 field campaign. Some samples are given in the attached powerpoint. Earth Observation (EO) Data Received/Used For each Mission/sensor: Landsat5 Space agency or Supplier: NASA Optical Number of scenes: 5 Range of dates: 08/09/2013; 11/11/2013; 30/01/2014;11/09/2014;29/10/2014 Beam modes/ incidence angles/ spatial resolutions:, 30 m MS + 15 m PAN Processing level: TOA reflectance Challenges, if any, in ordering and acquiring the data Challenges, if any, in processing and using the data Please provide sample pictures of the imagery provided in application: see below For each Mission/sensor: DEIMOS Space agency or Supplier: deimos imaging Optical Number of scenes: 3 Range of dates: 13/11/2013;05/04/2014;19/07/2014; Beam modes/ incidence angles/ spatial resolutions:, 20 m Processing level: TOA reflectance Challenges, if any, in ordering and acquiring the data Challenges, if any, in processing and using the data Please provide sample pictures of the imagery provided in application: see below

3 Landsat Landsat Deimos Landsat Deimos Deimos

4 Landsat Landsat

5 In situ Data Describe the in situ data collected, with methods and challenges, if any. We collected 847 GPS point in the field in December 2014, following the JECAM protocol and updated nomenclature for our site specificities. GPS points were chosen along roads to cover most part of the JECAM area (see figure below). GPS points were afterwards converted to polygons based on the images. The minimum homogeneous polygon of the 8 images described above was chosen. Illustration of the 847 polygons of the classified area The number of polygons for each class is given in the following table: Banana plantation 3 Built-up 53

6 Coffee plantation 14 Corn 30 Eucalypts plantation 160 Fallow 7 Forest 36 Orange tree plantation 63 Other 30 Pasture 127 Pines plantation 47 Rocks 11 Soybean 91 Sugarcane 154 Water 21 For communication purposes, please provide photographs of site work if available. Collaboration Have you been approached to participate in a collaborative project with other sites? Y/N If yes, please describe the nature of the collaboration ( i.e. Who, objective, brief status). SIGMA - JECAM experiment on medium to large field size agrosystems The main objective of this project is to test and compare classification methods for cropland area estimations based on MODIS data, and applied in different contrasted sites. These sites were selected within JECAM for their large field size agrosystems. The nature of the collaboration for the Brazil-SP site relies on data preparation and share, field expertise, review of the results obtained in this experiment, complementary measurements if necessary, etc. Results Describe your key the results, positive and negative. We used the 8 images described above to produce a land cover map of our JECAM site for December I give below a brief description of the method we used. First, a polygon segmentation of the images was performed under Trimble ecognition software. Then, 240

7 variables were computed for each polygon: we used all the bands reflectances of all images, and computed several vegetation indices. The randomforest algorithm was then used under R. The model was calibrated on the field data, and afterward applied on the entire image, giving the final landcover map. Some interesting outputs were also computed, as the classification stability, based on the membership probability. A confusion matrix is given below. The result is very good for sugarcane, eucalyptus, pines, forests, pastures and water bodies. Classification error is high for coffee plantations, maize and orange tree orchards. Banana Buildup surface Coffee Eucalyptus Maize Natural forest Orange tree Other Pasture Pinnus Soya beans Banana % Build-up % surface Coffee % Eucalyptus % Maize % Natural % forest Orange % tree Other % Pasture % Pinnus % Soya beans Sugar cane Water bodies Young fallow Sugar cane Water bodies Young fallow % % % % class error The classified surfaces of the JECAM site for December 2014 are given in the table below: Class area (ha) % of the total area Banana % Build-up surface % Coffee % Eucalyptus % Maize % Natural forest % Orange tree % Other % Pasture %

8 Pines % Rocks % Soybeans % Sugarcane % Water bodies % Young fallow % Some results are still unrealistic, like the coffee plantation, orange trees and maize area, but it is possible than next field campaigns will help to constrain more these classes. For communication purposes, please provide some graphic representation(s) of the results.

9 December 2014 classification

10 December 2014 classification stability (better when close to 1). This reflects more or less the fact that cropland (South-West and East) are classified with less membership probability

11 To what extent have the project objectives been met? The objectives have been met for the moment. We will continue to improve the map by adding new training polygons by photointerpretation, and with new field campaign. The method seems well adapted for the main class, but we got difficulties for class with small number of samples, as expected. One issue is the other class, which groups all land use that were met only several times in the field visit. The question of adding this other class in the classification algorithm is still controversial. One major difficulty is cropland classes since there is not a common regional calendar (e.g some soybean fields are just sowed when others are harvested). Therefore, we will increase the frequency of field visit in 2015, with measurements expected every 2.5 months. The Random Forest method needs also further explorations when images are partly covered by clouds. Use of textural variables computed on the panchromatic channels may also enhance some results, especially in row-structured fields such as orange tree orchards. While the algorithm is in theory able to handle a large number of variable, it could be usefull to keep only the more important ones (e.g. through a PCA analysis for instance). Can this approach be called best practice? Even if some aspects could be enhanced (see above), this method seems reliable enough. More investigations for improving the classification and the validation will be necessary Have you modified the project objectives? If so, in what way? Plans for Next Growing Season Next growing season, will you maintain your current approach, or modify the approach? If you plan to modify, please describe your new approach. We will maintain the current approach, by doing field inventories every months on the south-western part of the JECAM area, mainly covered by croplands. Images acquisitions are also planned this year (in particular SPOT images) Do you anticipate ordering the same type/quantity of EO data next year? Y, included DEIMOS data if possible If not, what type and quantity of EO data do you plan to acquire? We could acquire radar images if other JECAM partners are interested Publications

12 Please list any publications from your JECAM related research since last year s report (presentations, peer reviewed papers, technical reports, etc). le Maire, G., Dupuy, S., Nouvellon, Y., Loos, R.A., & Hakamada, R. (2014). Mapping shortrotation plantations at regional scale using MODIS time series: Case of eucalypt plantations in Brazil. Remote Sensing of Environment, 152,